Sources, attribution, and delivery profiles.

How the 96.09% continental coverage figure is defined

The continental coverage figure measures the road scope targeted by Sherpa across North America and Europe. Driveways, private access roads, and dead-end spurs outside the product scope are not included. The finished production layer contains 66.40 million OSM-way-keyed road-section classifications.

OpenStreetMap

OpenStreetMap supplies road geometry, topology, names, known surface examples, community-contributed road attributes, and selected route ingredients. Sherpa uses mapped surface examples as supervision, then reclassifies the full production network with newer imagery and the wider feature system. For roads with a usable source tag, the final Sherpa classification comes from a fold-specific model trained on neither that road nor its held-out geographic group. Roads without usable tags receive full-model inference. The finished layer is not a repackaging of OSM surface tags.

For the public comparison, a usable OSM surface tag means a value that the audited label-normalization rules could map consistently to paved or unpaved. On the same 66.40 million North American and European production road sections, 25.08 million met that definition, or 37.8%. This is a source-data coverage measurement, not the share of final classifications copied from OSM. Sherpa reclassifies all 66.40 million sections, including those 25.08 million tagged roads; 41.32 million are classified without any usable OSM surface tag. The selected Louisville-area visual uses the same denominator on both sides: 4,486 of 82,342 road sections had a usable source tag, while Sherpa independently classifies all 82,342.

OpenStreetMap is licensed under ODbL 1.0. Interactive maps credit © OpenStreetMap contributors. Customer packages are reviewed for the attribution and ODbL obligations applicable to their exact geometry, attributes, and format.

Overture Maps

Overture transportation, buildings/building parts, land use, land cover, water, places, addresses, and divisions provide geographic context where present. Structures and available height/floor information also support visibility and occlusion analysis.

Overture is a multi-license dataset. Several current themes are ODbL, while Places and Addresses may carry feature- or source-specific terms. Sherpa preserves release and source metadata and builds delivery-specific attribution from the official Overture attribution matrix.

Copernicus Sentinel-2 Global Mosaic

Surface reflectance in RGB, B04 red, and B08 near-infrared forms is converted into road-centered imagery, spectral measurements, learned visual representations, and supporting water/vegetation context. The inspected source mosaics span acquisition periods rather than one uniform “current” date.

Sentinel and Copernicus Land Monitoring Service information is available for commercial adaptation with acknowledgement. Learn more from the S2GM product page and CLMS use terms.

Harmonized nighttime light derived from VIIRS

A harmonized 2024 nighttime-light layer helps identify likely centers of human activity for modeled road-use exposure and supplies contextual measurements elsewhere in the system. It is not observed traffic, census population, or athlete/customer activity.

The source is Li, Zhou, Zhao, and Zhao's CC BY 4.0 harmonized nighttime-light dataset, version 10, DOI 10.6084/m9.figshare.9828827.v10. Deliveries carry attribution and a modification notice.

ESA WorldCover 2021

ESA WorldCover supplies a 10-meter land-cover prior that complements direct spectral evidence in scenic vegetation and landscape processing.

WorldCover 2021 is CC BY 4.0. Delivery and map credits follow the acknowledgement on the official data-access page and identify Sherpa's modifications.

Terrain and elevation

A combined 30-meter elevation surface supports road profiles, relief measurements, terrain context, and line-of-sight analysis. NASA SRTM is the principal source; Copernicus WorldDEM-30 supplies high-latitude coverage and gap filling. WorldDEM is a digital surface model, so structures and vegetation may be represented in the measured top surface.

NASA/JPL SRTMGL1 V003 is available without restrictions on reuse or redistribution, with citation requested. Modified WorldDEM derivatives are identified as produced using Copernicus WorldDEM-30 © DLR e.V. 2010-2014 and © Airbus Defence and Space GmbH 2014-2018 provided under COPERNICUS by the European Union and ESA; all rights reserved. The WorldDEM-30 license also requires its prescribed no-liability statement and downstream notice conditions.

Soil, climate, and socioeconomic context

ISRIC SoilGrids adds broad 250-meter soil texture and chemistry context. Long-term climate measurements have been evaluated as model inputs. Historical gridded GDP and HDI add regional development context, using the latest available 2015 band rather than a present-day local economic estimate.

SoilGrids 2.0 is CC BY 4.0. The Kummu GDP/HDI series is CC0 and spans 1990-2015; Sherpa nevertheless preserves scholarly attribution. WorldClim v2.1 requires permission for commercial use. A commercial model profile must either carry that permission or be retrained without WorldClim-derived inputs. Restricted raw rasters are not redistributed.

Vision and text model components

Sherpa's task-specific models combine road-centered image predictions with visual representations from DINOv2 and OpenCLIP, plus continuous text representations from BAAI BGE. The verified EfficientNetV2-S stage is fine-tuned for road-surface classification; these upstream models provide measurements rather than a final answer on their own.

BGE-base-en-v1.5 is MIT, and DINOv2 is Apache 2.0. The exact LAION OpenCLIP checkpoint is marked MIT, but its model card places deployed use outside its intended scope and calls for in-domain evaluation. The local pipeline does not use OpenAI's pretrained CLIP weights. timm's maintainers also flag ImageNet-origin diligence for pretrained weights. Sherpa therefore excludes upstream checkpoints and raw model embeddings from standard data deliveries and releases final predictions only through an approved, task-validated source/model profile.

Open geospatial, modeling, and routing software

H3 supports spatial grouping and context. RoutingKit contraction hierarchies accelerate continental shortest-path queries. CatBoost reconciles the structured model inputs. Leaflet 1.9.4 renders the interactive web map.

Software notices are retained when source or binaries are distributed; dataset outputs are packaged separately from software licensing. RoutingKit and Leaflet are BSD 2-Clause. H3 and CatBoost are Apache 2.0. Leaflet's notice is included with the deployed static assets.

Presentation imagery and optional basemaps

The portfolio layer samples use Sherpa's processed Sentinel-2 source mosaic, the same source family used in the imagery pipeline. The selected Louisville-area comparison defaults to an OpenStreetMap basemap for legibility and also offers a matched Sentinel-2 satellite view. The live atlas may offer optional third-party basemaps solely to help visitors orient themselves.

Optional live-atlas basemaps remain subject to their providers' terms and visible attribution. They are not model inputs, Sherpa-owned imagery, or part of a customer data delivery. No third-party basemap tiles or stitched basemap imagery are included with the licensable Sherpa road attributes.

What accompanies a delivery

• field definitions, score/support semantics, and release notes;
• source versions, compact and full attributions, license links, and modification notices;
• a provenance manifest appropriate to the selected fields and format;
• a clear distinction between upstream geometry/attributes and Sherpa-generated predictions, rankings, confidence, and diagnostics;
• the license-cleared source/model profile used for the delivery; and
• an explicit inventory of upstream rasters, pretrained weights, and raw embeddings excluded from the package unless separately cleared.